The motivation behind the study rests on a growing body of previous research suggesting that culturally mismatched interventions may provoke resistance in participants and reduce the effectiveness of music-based treatments. A student in China listening to a guided relaxation script written with Western references, for example, might feel disconnected from the imagery or simply fail to engage with it. Large language models, with their ability to produce fluent, contextually sensitive text on demand, present an obvious opportunity: therapists and researchers could generate verbal suggestions that resonate with a specific audience without the time and expense of manual adaptation. Whether such machine-generated content actually helps, however, had not been tested in a randomized design.

To find out, the team recruited 113 Chinese undergraduate students and randomly assigned them to one of three multicomponent conditions. The first group listened to Chinese film music accompanied by LLM-generated verbal suggestions that invoked Eastern imagery, drawing on cultural touchstones intended to feel familiar and soothing to the participants. The second group received a standardized music therapy recording with conventional verbal guidance, representing the existing gold standard of pre-packaged therapeutic audio. The third group listened to piano music without any verbal guidance at all, providing a baseline against which the value of spoken suggestions could be roughly gauged. All three conditions were delivered after participants completed a 40-minute cognitively demanding task, a standard laboratory method for inducing measurable stress and fatigue.

The researchers measured outcomes before and after the intervention across three domains. State anxiety, the transient feeling of nervousness and tension, was assessed with validated self-report instruments. Positive and negative affect, capturing the balance of pleasant and unpleasant emotional states, were measured separately, reflecting the well-established psychological principle that these are not simply opposite ends of a single scale. In addition to subjective reports, the team recorded heart rate, offering an objective physiological indicator of arousal and relaxation. This combination of self-report and biological measurement is a strength of the design, since subjective and physiological responses to music do not always move in tandem.

The statistical analysis used baseline-adjusted analyses of covariance, or ANCOVAs, which control for participants’ starting levels on each measure and thereby sharpen the ability to detect genuine change attributable to the intervention. On positive affect, the three groups did not differ significantly, with the analysis yielding F(2, 109) = 1.03, p = .360, and a small effect size of ηp² = .019. State anxiety likewise showed no group differences, F(2, 109) = 0.68, p = .511, ηp² = .012. In other words, none of the three conditions, including the culturally adapted LLM-enhanced one, produced a detectable advantage in boosting positive emotions or reducing self-reported anxiety compared with the alternatives.

The picture changed, however, for negative affect. Here the analysis revealed a significant group effect, F(2, 109) = 4.60, p = .012, with a moderate effect size of ηp² = .078. Post-hoc comparisons with Holm adjustment for multiple comparisons showed that the Eastern imagery group had lower adjusted post-intervention scores than both comparison groups, with adjusted p values of .026 for each contrast. This suggests that students who heard the LLM-generated, culturally tailored suggestions alongside Chinese film music reported a greater reduction in unpleasant emotional states than those who received the standardized therapy recording or the piano music alone. For a pilot study, a moderate effect size on any outcome is a signal worth pursuing.

Heart rate told a different story. This physiological measure also differed significantly across groups, F(2, 109) = 10.56, p < .001, and with a notably large effect size of ηp² = .162. But the pattern ran counter to the subjective findings: it was the standardized music therapy group, not the Eastern imagery group, that showed lower adjusted heart rate values than the other two conditions. This dissociation between what participants felt and what their bodies were doing is a well-documented phenomenon in stress research and underscores why the authors were careful not to overinterpret their results. A culturally resonant narrative may soothe the mind while the music’s acoustic properties, tempo, and familiarity drive the body’s response, and the two pathways need not align.

Indeed, the authors are explicit about the limits of what this study can establish. The findings identify potentially meaningful response patterns but do not demonstrate efficacy, nor do they isolate the independent effect of the culturally adapted verbal suggestions. The reason lies in the design: the three conditions differed not only in the presence or cultural content of the verbal guidance but also in the musical material itself and in the duration of the intervention. The Eastern imagery group heard Chinese film music, the standardized group heard a different therapeutic recording, and the piano group heard yet another piece. Any of these differences, alone or in combination, could account for the observed group effects. Disentangling the contribution of the LLM-generated suggestions from that of the music would require conditions that hold the musical stimulus constant while varying only the spoken content.

These caveats are standard and appropriate for a pilot randomized exploratory study, whose primary purpose is to test feasibility, refine procedures, and generate hypotheses rather than to deliver definitive answers. By that standard, the trial succeeds on several fronts. It demonstrated that LLM-generated culturally adapted suggestions can be produced and integrated into a music-based intervention in a way that participants completed without apparent problems. It registered the trial publicly on ClinicalTrials.gov under identifier NCT07359950, signaling a commitment to transparency. And it produced at least one statistically significant, moderately sized effect on negative affect, along with a large physiological effect on heart rate, that together justify larger and more tightly controlled trials.

The broader implications extend beyond music therapy. As large language models move into health care, questions about cultural adaptation are becoming urgent, and this study offers one of the first randomized tests of machine-generated, culturally tailored therapeutic content in a real intervention. The results hint that AI-generated scripts rooted in familiar cultural imagery may help reduce negative emotional states in stressed students, even as the physiological evidence complicates the story. If future trials with larger samples, matched musical materials, and longer follow-up confirm and extend these patterns, the convergence of generative AI and culturally informed therapy could become a practical, scalable tool for supporting student mental health. For now, the study stands as a careful, honest first step, one that maps the terrain and marks the spots where the next, more definitive expeditions should dig.

Subject of Research: A pilot randomized study testing LLM-generated culturally adapted verbal suggestions integrated into music therapy for anxiety and affect in Chinese university students.

Article Title: Integrating large language model-generated culturally adapted verbal suggestions into music therapy for university students: A pilot randomized exploratory study

Article References: Li, W., Wang, G., Manoharan, P., Cui, X., Dai, L., & Huang, L. (2026). Integrating large language model-generated culturally adapted verbal suggestions into music therapy for university students: A pilot randomized exploratory study. PLOS One, 21(10), e0360478. https://doi.org/10.1371/journal.pone.0360478

Image Credits: AI Generated

DOI: 10.1371/journal.pone.0360478

Keywords: music therapy, large language models, cultural adaptation, state anxiety, negative affect, heart rate, randomized controlled trial, university students, mental health, PLOS One, guided imagery, pilot study